Supporting medical decision-making using machine learning

WAINWRIGHT, Richard (2023). Supporting medical decision-making using machine learning. Doctoral, Sheffield Hallam University.

Preview

PDF Wainwright_2023_PhD_SupportingMedicalDecision.pdf - Accepted Version Creative Commons Attribution Non-commercial No Derivatives. Download (24MB) | Preview

Abstract

As the strain on health care continues to grow worldwide, the need for reliable decision-making has never been more apparent. The computerisation of electronic health records has provided a wealth of data that can be applied to various medical use cases. Machine Learning algorithms are exploited to try and assist with making effective decisions. The resulting contributions within this work demonstrate that it is possible to lean on advancements in computer science to develop support tools for medical practitioners which assist in their decision-making processes. This thesis contributes four core advances to the research domain: Firstly the enhancement of current mortality prediction systems in intensive care units was considered. Comparing multiple Machine Learning classifiers with optimised pipelines produced results that were both comparable and more effective at determining patient mortality than the existing APACHE II model. The most encouraging classifier was Decision Trees whilst being trained using: K-fold cross validation, Grid search hyper-parameter tuning and SMOTE achieving an average AUROC score of 0.93 and accuracy of 0.92. Unlike other mortality prediction systems which are often trained on small cohorts of data, a method of retraining and optimising for different patient cohorts is introduced. Retraining based on a patients age or admission in to the ICU is also considered as a novel approach of keeping support tools up to date. An ensemble imputation method has been developed that can be used to generate the missing data in a real life dataset. This has produced accuracy and recall results comparable to current state of the art techniques when applied to the Cleveland hospital dataset. In this work, strategies to rebalance datasets are investigated to predict early onset Sepsis. One promising approach examined in this thesis is the use of the RUSboost algorithm. This enabled the optimisation of a classifier that has a high fidelity without overfitting.

Item Type:	Thesis (Doctoral)
Contributors:	Thesis advisor - Shenfield, Alex [0000-0002-2931-8077]
Additional Information:	Director of studies: Dr. Alex Shenfield
Research Institute, Centre or Group - Does NOT include content added after October 2018:	Sheffield Hallam Doctoral Theses
Identification Number:	https://doi.org/10.7190/shu-thesis-00542
Depositing User:	Colin Knott
Date Deposited:	25 Aug 2023 16:31
Last Modified:	11 Oct 2023 12:15
URI:	https://shura.shu.ac.uk/id/eprint/32312

Actions (login required)

View Item

Downloads

Publication Metrics